Geographic variations and temporal trends of Salmonella-associated hospitalization in the U.S. elderly, 1991-2004: A time series analysis of the impact of HACCP regulation

Individual hospitalization records were obtained from the Centers of Medicare and Medicaid Services (CMS). The dataset contained age, residential state and zip code, date of admission, and up to 10 International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnostic codes of all the hospitalized Medicare recipients aged 65 or above in the U.S. from 1991 through 2004. The CMS data are nationally representative because Medicare covers about 95% of the U.S. elderly. In addition, hospitalizations due to Salmonella infections in the U.S. were shown to be significantly correlated with the surveillance records [3], hence are valid estimations of overall salmonellosis rates.

Salmonella infections are represented by two ICD codes which displayed very different incidences: In the 14-year duration, there were only 584 cases of ICD 002, typhoid and paratyphoid fever, but 28,959 cases of ICD 003, other Salmonella infections. So, in this analysis, we adopted ICD 003 as the only condition and defined a hospitalization as Salmonella-associated if the first three digits of any one of the 10 diagnostic codes were "003". Under category 003 were different sub-codes representing various manifestations of the disease. For instance: 003.0 is gastroenteritis, 003.1 is septicaemia, and 003.21 is meningitis. Serotypes, however, were unavailable in this dataset.

Data aggregations were performed prior to analysis. The individual records were aggregated from daily level (5,114 days) into weekly level (734 weeks), as well as from the county level into Census divisions, which is a geographic categorization used by the Census Bureau that divides the U.S. into nine areas; New England, Middle Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, and Pacific [17]. Eight of the Census divisions are composed of contiguous states. The exception, Pacific, includes the states Hawaii (HI) and Alaska (AK). Since Salmonella infection is associated with environmental factors such as temperature [15, 18], we decided to exclude Hawaii and Alaska from this analysis to maintain the homogeneity of the potential exposures in the Pacific division.

The major outcome was the hospitalizations rate, derived using the population data from Census 1990 and 2000. Weekly divisional elderly populations were estimated with linear interpolation.

This analytical protocol was approved by the Institutional Review Board at Tufts Medical Center/Tufts University Health Sciences Campus.

A linear spline model (also known as broken stick model) was used to discern if the slope changed significantly before and after the start of 1997--when the first stage of the HACCP system on bacterial testing in broilers started.

Long-term trends and accompanying seasonality were assessed with the seasonal-trend decomposition procedure based on LOESS [19], which assumes that a seasonal pattern is an additive result from three components: trend, seasonality, and remainder. Based on this idea, we modified and applied a Poisson regression model to detect the changes of long term trends before and after HACCP introduction while controlling for seasonal fluctuation:

where Y _t is a time series of the weekly hospitalization rates modelled with Poisson distribution. The linear terms, t _<1997 and t _{≥ 1997} represent the time (in weeks) before and after the HACCP implementation (beginning of 1997), they were constructed so that:

allowing the long term trend of hospitalizations in the pre- and post-HACCP period regression line to inflect at the first week of 1997. The intercept, β ₀, represents the predicted outcome at the inflection point. Through comparing the confidence intervals of the coefficients β ₁ and β ₂, significant change in long-term trend before and after the HACCP implementation can be inferred.

The two trigonometric functions--sin(2πωt) and cos(2πωt)--serve to adjust for annual oscillations of the outcome; ω is the frequency of an annual oscillation expressed in weeks and the constant 2π specifies a one-cycle annual oscillation.

In order to test whether the annual seasonal characteristics were different, we applied harmonic regression on every single year of the data, from 1991 through 2004, to generate a set of comparable estimates. The model is as follows,

where k represents each individual year and other notations are identical to that of formula (1). The estimates β ₁ and β ₂ in Formula (2) were used to derive the time to peak, minimum, maximum, and relative intensity (also known as peak-to-trough ratio, derived by dividing the maximum by the minimum). Details on this method were presented by Naumova and McNeill [20].

A generalized linear model with Poisson distribution and log link function was used in both linear spline and annual harmonic regression analyses. All data analyses were carried out with S-Plus 8 [21]. Statistical significance is declared based on α = 0.05. Mapping was performed with ArcGIS 2.0 [22]. All geographic analyses were performed to each Census division as well as the whole contiguous United States in order to allow comparisons against the average.

From about 200 million hospitalization records, we abstracted 27,790 records that involved Salmonella infection. Figure 1 shows the hospitalization counts in the 14-year duration, annual rates, and temporal fluctuation represented by sparkline (Loess curve, 5% span) [23] of hospitalizations related to Salmonella infection in the contiguous U.S. between 1991 through 2004.

The South Atlantic, East South Central, and West South Central divisions showed pronounced annual oscillations. Long-term decreases were observed in the New England, Middle Atlantic, Pacific, and Mountain divisions.

Table 1 shows the results of the linear spline analysis. The results were rearranged and displayed in Figure 2, in which divisions with similar behaviors in their change of hospitalization rates were grouped and described.

Figure 3 displays the time series plots, on which the fitted values of the long-term trend analysis were plotted, for all nine divisions as well as the contiguous U.S. The national map at the center provides average weekly rate information from 1991 to 2004. The vertical reference line indicates the beginning of 1997.

To demonstrate the annual variation, unadjusted between-year confidence intervals were computed (N = 14 years) and displayed in Table 2 for the time to peak, minimum and maximum rates, as well as relative intensity (peak-to-trough ratio). The linear term for the week series (t _k ) was found to be statistically insignificant in all the models and hence we removed this term from the model. The divisions are ranked by descending order of their annual rates retrievable from Figure 1.

The peak times of hospitalizations due to Salmonella infections ranged between 28^th to 35^th week of a year (about July till mid August). West South Central, East South Central, and South Atlantic had annual rates exceeding the national reference. They also had significant decrease before HACCP, but such trend became insignificant afterwards (Table 1).

It is important to realize that the raw difference in divisional rates is not the outcome of interest for two reasons: (i) the rate of hospitalization is not just a function of disease prevalence but also of other factors, such as the number of hospital beds per citizen, practices of disease coding, and health-seeking behaviors, (ii) the implementation of the HACCP regulation studied here was very unlikely to have impacted on the divisional medical infrastructure, in this analysis, we chose to focus on the change of hospitalization rates in the pre- and post-HACCP period within each division, and it is these changing patterns that we wish to compare and contrast.

The change of hospitalization rates before and after the HACCP was unequal across the divisions. Some divisions did not show expected results: the hospitalization rates in East South Central, West South Central, and South Atlantic divisions were originally decreasing before pre-HACCP, however, they became stable in the post-HACCP period (Figure 2). The East South Central, West South Central and South Atlantic divisions persistently displayed a higher hospitalization rate, estimated maximal hospitalization rate, and relative intensity (Table 2), while their rates of change in hospitalization shifted from decreasing in the pre-HACCP period to unchanged in the post-HACCP period.

Despite the above findings, we cannot nullify the potential beneficial effect of the HACCP regulation in the aforementioned divisions, because the data we used are inappropriate for an intervention-control analysis; it is possible that hospitalization rate would have been even higher if no regulation were introduced at all. In addition, referring to Table 2, the predicted rates of most Census divisions did drop; on average, the predicted minimal rates decreased in most of the divisions except the West South Central (p < 0.05), while the predicted maximal rates also decreased significantly (p < 0.05), except in East South Central. Only the West South Central showed a significant decrease of relative intensity (from 4.22 to 3.26, p < 0.05).

To further discern the possible causes of the patterns observed, investigations involving the collection of data on microbial contamination and detailed bacterial identification, such as serotypes, should be adopted. Nevertheless, the stable rates in three divisions provide important information for geographic targeting.

Geographically, the divisions with higher hospitalization rates are mostly situated in the south-eastern United States, where several features might have contributed to these high rates. First, the general health of the elderly in these areas could be relatively worse: in 2000, the all-cause hospitalization rates of elderly aged ≥ 65 in the East South Central and West South Central divisions were 39.7% and 34.4%, respectively, compared to the national average 29.3%. Second, the temperature and humidity in this area tend to be higher [24], thus favouring the growth of thermophiles like Salmonella. This argument is further supported by the synchronized and similar temporal patterns of temperature and hospitalizations [15, 18, 25]. Third, according to Census 1990, sources of drinking water in the area were more likely to be surface water, which is more prone to contamination. Fourth, these divisions also have the highest density of broiler chicken industry in the U.S. [11], possibly leading to increased exposure to Salmonella in chicken waste. Fifth, the health-seeking behavior in these mostly rural divisions might also confound with the hospitalization outcome, since distance from the closest hospital was found to influence the rates of hospitalization [26]. Finally, the area is also prone to Atlantic storms and hurricanes, which can cause floods and destroy infrastructures that indirectly lower the quality of the surface water and overwhelm the water treatment systems [27].

Using the West South Central division as an example, we found that the times to peak were overall similar to the national one, which is about the 25^th to 31^st week (late June to early August) in a year, with a slight increasing trend throughout the 14 years. The relative intensity is consistently higher compared to the national reference. The combination of high rate, moderate counts (13% of hospitalizations related to Salmonella infections were in this division), and consistent time to peak indicates this division to be a potential target for preventive programs against Salmonella infections.

As time and resources to achieve the Healthy People 2010 goal appear limited, it is important to carry out accurate spatial and temporal targeting for maximum return of benefit. We have identified the South Atlantic, East South Central, and West South Central divisions to be good potential geographic targets, and the 33^rd week to be the predicted peak of the rate of hospitalizations due to Salmonella infections. Preventive programs, such as water quality monitoring, public service announcement, and other related educational components, can be administered at the most optimal times and places. Of particular concern are the institutionalized elderly, who tend to have longer stays in the hospital and higher mortality rates [28].

There are, however, some limitations to this study. First, we acknowledge that the criteria of coding a disease as Salmonella infection could have changed in the study period. However, ICD 003 has been introduced since the Eighth Revision in 1965. In addition, according to National Center for Health Statistics, this code has not been revised or reassigned within the period of our study [29]. We can safely assume that the practice of coding a case of salmonellosis is well established, at least within each division.

Second, there is no validity study on the coding of this particular disease. However, Dubberke et. al. showed a satisfactory agreement between the ICD codes and individual laboratory test results in Clostridium difficile, one of the major gastrointestinal infections among the elderly [30]. Studies with similar settings using ICD-9-CM as an outcome of interest were done [31, 32]. This analysis provides further insight into applying time series analysis to the national medical records with the purpose of targeting service recipients in both geographic and temporal means. It is also one of the first studies evaluating the potential impact of HACCP implementation on the U.S. elderly.

Fourth, we assumed that the HACCP started at the first week across the nation and the degree of compliance was the same across the division due to the lack of such desirable data. Due to the same reason, we could not study association in different types of meat production facilities. However, chickens were a major transmission route because they have high Salmonella prevalence and are preferred by the Americans. According to the USDA, the baseline prevalence rates of testing positive with Salmonella in ground chicken and broilers were 45% and 20%, respectively, much higher compared to hog (9%) and beef (3%) [33]. Also, as of 1992, per capita consumption of chicken exceeded that of cattle [34]. We can safely assume that any major improvement brought about by HACCP probably happened through the broiler chicken and ground chicken industries.

Lastly, the CMS dataset, despite its high coverage, is susceptible to under-testing. Also, different health-seeking behaviors might have affected the number of hospitalizations. Up to 92% of all the records in this analysis were "apparent" one-time hospitalizations. Due to the different deidentification processes in each batch of data, there was no valid method to completely identify repeated hospitalizations.